What is a deepfake?

How it works

The US Government Accountability Office (GAO) described a deepfake in February 2020 as a video, photo or audio recording that appears real but has been altered using AI. It outlined two techniques built on artificial neural networks, software loosely modeled on the brain: autoencoders and generative adversarial networks (GANs). Autoencoders learn to rebuild an input from a simpler representation. In GANs, one network produces fakes and another tries to detect them.

The word was coined in late 2017 by a Reddit user of the same name, according to MIT Sloan. In EU law, Article 3 of the AI Act covers image, audio or video content that:

• is generated or manipulated by AI,
• resembles existing people, objects, places, entities or events, and
• would falsely appear authentic or truthful.

Why it matters

The FBI warned on December 3, 2024, that criminals use AI-generated audio to imitate relatives in fake emergencies and AI-generated video to pose as executives on live calls. Its 2025 IC3 Annual Report counted 22,364 complaints that referenced AI, with adjusted losses above $893 million. That category is wider than deepfakes alone.

A March 2024 GAO review said existing detection methods may not accurately identify deepfakes in real-world conditions. It said a fake can spread disinformation as soon as it is viewed, even if it is later exposed. A November 2024 NIST report reviews other approaches, including provenance tracking, watermarking and labeling.

Where things stand in 2026

In the US, the TAKE IT DOWN Act became law on May 19, 2025. It makes it a crime to knowingly publish nonconsensual intimate images, including AI-made ones it calls digital forgeries. It gave covered platforms until May 19, 2026, to set up a process for removing such content within 48 hours of a valid request.

In the EU, the AI Act's transparency obligations have applied since August 2, 2026, according to the European Commission. Providers must mark AI-generated output in a machine-readable format. Deployers must disclose deepfakes. The Commission's implementation timeline gives certain providers with systems already on the market until December 2, 2026, to meet the marking rule. It lists the same date for new prohibitions on AI systems that generate non-consensual sexual deepfakes or child sexual abuse material. A voluntary code of practice was published on June 10, 2026. The Commission counted about 190 signatories by the end of July.